Guaranteed-Quality

Mesh Generation

for Curved Surfaces *

L. Paul

Chew

Department of Computer Science Cornell University Ithaca, NY 14853

ABSTRACT
For several commonly-used solution techniques for partial

dhlerential equations, the first step is to divide the problem region into simply-shaped elements, creating a mesh. We present a technique for creating high-quality triangular meshes for regions on curved surfaces. This technique is an extension of previous methods we developed for regions in
the plane. For both flat and curved surfaces, the resulting meshes are guaranteed to exhibit the following properties: (1) internal and external boundaries are respected, (2) element shapes are guaranteed - all elements are triangles with angles between 30 and 120 degrees (with the exception of badly shaped elements that may be required by the specified boundary), and (3) element density can decontrolled, producing small elementsin interesting areas andlargeelements elsewhere, An additional contribution of this paper is the development of a practical generalization of Delaunay triangulation to curved surfaces, 1

2. We extend this meshing technique to curved surfaces. Just as in the 2D meshing technique, triangles in the meshes produced by the curved-surface technique are guaranteed to respect boundaries and to be both wellshaped and well-sized. For curved surfaces, a user is likely to want welLsized to include the criterion that a triangle is a good approximation to the corresponding surface. 3. We develop a practical definition of Delaunay trian-

gulation for curved surfaces. Both the 2D and the curved-surface techniques are baaed on Delaunay triangulations. There is a large literature on the subject eration with most of the material emanating placations community. One good information interested reader is the recent survey by Bern of mesh genfrom the ap source for the and Eppstein

INTRODUCTION
three main contributions to the subject

This paper includes of mesh generation:

1. We show that a relatively simple mesh generating technique, based on earlier work by the author [Che89], can be used to create meshes in which (1) region boundaries are respected (including internal boundaries) and (2) all triangles are guaranteed to be both well-shaped and ruell-sized. Well-shaped triangles have angles that are between 30 and 120 degrees (smaller angles may appear if small angles appear as part of the boundary). Well-sized is defined by the user and can mean virtually anything as long w it can be achieved by shrinking the triangles. Typically, a triangle might be considered well-sized if its estimated error is small based on the users intuition or on an earlier, presumably-cruder, numerical solution.
*This Agency 92-J-1989, IRI-9006137, work of the and was supported Department of by ONR AFOSR and by the Advanced Defense under ONR NOO014-92-J-1839, AFOSR-91-0328. Research Contract NSF Projects Noo014 Contract

[BE92]; th~ paper aleo includes a very good bibliography. See [BEG90, MS92, MV92, Rup92] for recent work related to our results. We assume the reader is familiar with Delaunay triangulations and with the incremental algorithm for producing them. We use a variation of the Delaunay triangulation called the constrained Delaunay triangulation (CDT). Bssically, the CDT, like the Delaunay triangulation, can be buflt by edge-fllpping, but for the CDT there are certain edges that cannot be filpped. The CDT is useful for mesh generation since it respects boundaries as well as having many of the properties of the Delaunay triangulation. The necessary background material is contained in, for instance, [BE92]. We develop the 2D meshing algorithm and prove some of its properties in Section 2. In Section 3 we describe the curved-surface Delaunay triangulation, which is used in Section 4 to create an algorithm for meshing regions on curved surfaces. The final section discuss practical improvements to the algorithms and some conclusions. 2

MESHING

IN THE

PLANE

Contract Contract

Permission granted direct that

to

copy that

without

fee

all or part are not made

of this

material

is for

provided commercial

the copies

or distributed notice

We azsume that our region-to-be-meshed has boundaries that consist of line segments (curved boundaries can be handled as well, but complicate the presentation). We use an implementation of CDT that supports the following operations:
q

sdvantage,

the ACM

copyright

and the

Build: tices),

given build

an initial the CDT

set of sources for this set. into

(edges

and ver-

title of the publication and its dfIte appear, and notioe is given
copying specific ACM is by permission To copy otherwise, permission. Geometry,5/93/CA, -583 -619310005 /0274 USA . ..$1 .!50 0-89791 of the Association or to republish, for Computing requires a fea
q

Machinery. and/or 9th W 1993 Annual

Insert: insert the given vertex date the CDT accordingly.

the CDT

and up

Computational

274

Figure

1: A mesh created

using

a size function

to control

element

density.

Delete: delete the given vertex from the CDT and update the CDT accordingly. (Note: endpoints of sourceedges are never deleted.) Split: given a source-edge, split it in half by adding a new vertex and replacing the source edge by two new edges. In addition, we assume that Split has an option where the given source edge can be split into thirds instead of halves.

b Travek Given a source vertex s (Part of the CDT) and a destination vertex d (not necessarily part of the CDT), determine if s and d can see each c,ther (i.e., determine if the segment from s to d intersects a source-edge at a non-endpoint). This is implemented by walking across the CDT from triangle to triangle. The following information is returned: If .s and d can see each other then return angle of the CDT that contains d; the tri-

For example, the user could specify a size function over the entire problem region; a triangle is well-sized if its circumradius is less than or equal to the value of the size function evaluated at the circumventer. Figure 1 shows an example of a mesh generated using such a size function. This kind of function works as long as there is lower bound 8>0 on the sizes that can appear within the problem region. The user may want to use a size function derived from. an error analysis of an earlier, presumably cruder, numerical solution to the problem. Another possible user-choice for well-sized triangles is to leave interior triangles completely unconstrained, but to require triangles with boundary edges to have boundary edges smaller than a chosen constant. (1) reqpired vertices: There are two kinds of vertices: those that are part of the boundary or those that are specifically chosen by the user, and (2) circumventer vertices: additional vertices that are introduced during the meshing algorithm. During the meshing process it is sometimes necessary to eliminate vertices. Note that only circumventer vertices can be eliminated; the other vertices must be retained. To show that the following algorithm does not keep meshing forever, making smaller and smaller triangles, we show that no edge introduced by the algorithm will have length less than h, where h is defined as the minimum among (1) the smallest line-of-sight clistance between two nonintersecting sources (vertices or edges) in the initial set of sources (note that objects on opposite sides of a source edge cannot see each other aud are thus infinitely far apart), (2) one half the length of the smallest source edge that appears in
the initial set of sources, and (3) the value $ associated with

Otherwise return the source-edge that blocks from seeing d (i.e., the first source edge that intersected on the way from s toward d). During the meshing cording to two criteria: algorithm, triangles shape and size.

s is

are gradec~ ac-

if all its anglesare Definition 1 A triangle is well-shaped greater than or equal to 30 degrees. A triangle is well-sized if it satisfies a user-supplied grading function. This function at ail as long as there exists can use any user- choaev criteria a value 6 > 0 such that any well-shaped triangle that fits within a circle of radius 8 would satisfy the grading function.

the user-supplied size grading function such that any well-shaped triangle that

(by definition, 6 is fits within a circle

275

of radius 6 is also a well-sized triangle). The quantity h is used in the algorithm only in step 6. To simplify our presentation, we assume that our initial problem region contains no angles that are less than 60 degrees. In fact, angles down to 30 degrees can be handled without significant modification to the algorithm. Angles less than 30 degrees can be removed from the problem region and triangulated in a postprocessing step; this type of lopping off technique is used in [BEG90, MS92, Rup92]

small /

region

ALGORITHM
1.

A
for the initial set of sources (edges and

Build the CDT vertices).

2. Grade

any triangles that are currently ungraded. A triangle passes only if it is both (1) well-shaped: its smallest angle is greater than 30 degrees, and (2) wellsized: it satisfies a user-defined grading function as described in Definition 1. Halt. Otherwise choose (call it A) and determine the its

Figure vertex

2: When e has length less than must lie within this small region.

fib,

the triangles

3. If all triangles

pass then largest triangle that fails circumventer (call it c). across the we either containing

4. Travel
c until triangle

CDT from any vertex run into a source-edge c. c then

of A toward or find the

Proo~: The proof is by induction. Certainly the lemma holds initially. We assume the lemma holds at cycle i and show that the lemma must continue to hold for the next cycle. We need to check steps 5 and 6, the only steps where new vertices are int reduced, to see that the lemma continues to hold. In Step 5, the circumventer of a failing triangle A is inserted into the CDT. Note that by definition of the CDT, As circumcircle must be empty; thus, the new vertex is no closer to any other vertex than the radius of this circumcircle. We claim that this radius is greater than or equal to h. To see thw, note that triangle A must have failed during the grading process. There are 2 ways for a triangle to fail: 1. Triangle A has a small angle. By simple geometry, a small angle has a small side opposite it; in fact, an angle is less than 30 degrees if the opposite side has length less than the circumcircle radius. Thus, when there is a small angle, the radius is larger than the length of some edge that already exists in the CDT, which by the induction hypothesis has length greater than or equal to h.

5. If we found
the CDT.

the triangle containing Go to step 2.

Insert

c into

6. If we ran into a source-edge

then Split the edge and update the CDT. This is always a one-half split unless the length of the edge is between 2fih and 4h, in which case we use a one-third split. Let / be the length of the new edges and consider the one or two new vertices produced by the Split operation. Delete each circumventer-vertex of the CDT that is closer than 1 (line-of-sight distance: an intervening source edge implies infinite distance) to a new vertex. Go to step 2.

In step 3, largest triangle means the triangle with the largest circumcircle. We could actually choose any failing triangle rather than the largest and the algorithm will produce a reasonable mesh. It might even take a bit less time, since choosing the largest requires maintaining a priority queue of some sort. However, experience has shown that the resulting mesh is more pleasing to the eye if larger triangles are done first. It would be useful to discover how well pleasing to
the eye correlates with better for finite element analysis.

2. Triamzle

It does not matter which triangle vertex we start at in Step 4, since the circumventer we are looking for is the center of an empty circle. (Note: by definition, each triangle of the CDT has a circumcircle that is emptg in the sense that no other vertices can be seen from the interior of the triangle.) Thus, during the Travel operation, we cross only edges whose endpoints are outside the circumcircle. The one-third Split in Step 6 is used to significantly simplify the proof of Lemma 2. A more complicated proof can be used to show that such one-third splits are actually unnecessary (although some smaller edges might be produced). Also in Step 6, nearby vertices are deleted in order to avoid introducing new edges that are short. This too is important in the proof of Lemma 2. Lemma 2 At Step ,2 of A/gorithm A, have line-of-sight distance less than h. no two vertices ever

A is not well-sized accordirm to the users size-g~ading function. To produce an-edge of length less than h, the radius of As circumcircle must be less than h. But when the circumcircle is less than h, A is automatically well-sized due to the definition of well-sized and the way h was chosen.

is split into either two or three In Step 6, a source-edge pieces and nearby vertices are eliminated. A source edge e is split only when there is a triangle on one side of e with the triangles circumventer on the other side of e. Assume that splitting e produces new edges that are too small. Let d be the length of e. Length d must be between h (by the induction hypothesis) and 2h. We claim that if d is less than then no splitting can occur. Too see th~, observe that, with e so short, the locations of the vertices of triangle A, the triangle that caused the split, are very constrained. The endpoints of e are possible vertices of A. Any other vertex (or vertices) of A must (1) lie in a region within or on the circle that has e. as its diameter (this must be true to make the circumventer fall on the correct side of e) and (2) must lie outside circles of radius h about each endpoint of e by the induction hypothesis. fih

276

The resulting small region (see Figure 2) is too small to fit more than single vertex; more vertices in the region would contradict the induction hypothesis. Thus, the vertices of A are the endpoints of e plus an additional vertex u that lies within the small region described above. This small region is so small, in fact, that all the possible triangles that use v and the endpoints of e are well-shaped and have circumcircles with radius less than h; thus, by the definition of well-sized, the triangles are also welI-sized. Any such triangle would pass in our grading step and could not possibly lead to edge split ting. At this point, we have shown that our source-edge e, that is assumed to split into edges that are too small, must have length between fih and 2h. We refer to an edge with length in this range as a bad edge. We claim that bad edges can never occur. To see this, we first note that, by the way h was defined, we initially start with all edges having length greater than or equal to 2h. Thus, a bad edge appears only if it is the result of an earlier Split operation, either a halfsplit or a third-split. To be the result of a half-split, the unsplit edge must have had length between 2fih and 4h, but edges with length in this range are never split in halfi such edges are split into thirds, producing edges with lengths in the range 1.1547h and ~h (outside the bad range and still of length greater than h). To be the result of a third-split, the unsplit edge must have had length between 3fih and 6h, but edges with lengths in this range are never split into thirds. Thus, bad edges never occur, so new source-edges can never have length less than h. We still need to show that the new splitting-vertex is not too close to other vertices ofour mesh. This is taken care of by the the elimination of nearby vertices that is done in. the remainder of step 6. Our definition of h and our restriction that no angles less than 60 degrees occur in our original region ensure that required vertices cannot be too close. u This lemma is just what we need to show that Algorithm A halts. Note that each cycle of the algorithm introduces either a new vertex along the boundary (which may eliminate some vertices in the interior) or a new vertex in the interior. The boundary has finite length and adjacent vertices on the boundary can be no closer than h, so points are added to the boundarv . at most a finite number of times. The interior of the region has finite area and, since vertices are no closer

For variable-density meshis, in some sense, optimal. ing, all of the optimality results of which we are aware [BEG90, MS92, MV92, Rup92] prove their resulting mesh optimal within a constant factor in the sense that the number of triangles they produce is within a constant factor of the number of triangles in the best possible mesh that achieves the same triangle-quality. Unfortunately, the constant-factors in the proofs are extremely IIarge; thus, the optimality results are not very helpful in terms of choosing which algorithm produces the smallest number of triangles in practice. As in Rupperts work [Rup92], ,Algorithm A is derived from the techniques presented in [Che89]; thus, an optimality result for Algorithm A (or a slightly modified version of Algorithm A) can probably be derivecl using some of Rupperts techniques. In practice, it appears that Algorithm A does not produce many more triangles than those necessary to achieve the 30 degree bound and reach the users size demands. 3

CIRCLES

ON

CURVED

SURFACES

We wish to extend Algorithm A to deal with meshes on curved surfaces. Algorithm A used two bazic ideaz to create a mesh: (1) the mesh is a Delaunay triangulation and (2) new vertices should be circumventers of Delaunay triangles. To make these ideas work for curved surfaces we have to determine reasonable definitions for a Detaunay triangulation on a curved surface and for the circumventer of a trianzle on a curved surface. Both of these things come downto determining what is meant by a circle on a curved surface. One possible definition for a circle, perhaps the most obvious such definition. is to use geodesic distance alorw the surface. Under this plan, a circ~e about a given cen~er is the set of points on the surface that are equidistant (taking the shortest route along the surface) from th~e center. This has two major disadvantages: (1) for most surfaces, it is very dHficult to calculate geodesic distances and (2) odd behaviors are possible (for inst ante, consider a circle-center on the side of a very steep hill - the circle can reach all the way around the hill, intersecting itself, without including the top of the hill). It does not seem possible to completely eliminate this kind of odd behavior, but we can at least attack the first difficulty by using a different definition of circle. Definition 4 Given three vertices on a curusd surface, consider the infinite set of spheres through the three vertices. The centers of all the spheres lie on a single line. We choose the sphere whose center is on the surface and define the cirof the thee vertices to be the set ~Yfpoints where cumcircle this sphere intersects the surface. This
q

than h, there are only finitely many vertices that can be fit into the interior of the region. Thus, we have the following
theorem. Theorem 3 Algorithm A halts.

Since Algorithm A halts, all the triangles of the mesh it produces must be triangles that pass during the grading step. Thus meshes produced by Algorithm A have the following properties:
q

definition

has some advantages:

Boundaries and vertices specified by the user are respected. Triangles do not cross boundaries. IJserspecified vertices are never eliminated. All triangles have angles greater than 30 degrees. (Smaller angles are required if we allow boundary angles less than 30 degrees.) The user can control triangle size. Any kind of size criterion can be chosen as long as it can be achi,eved by making triangles smaller.
q

It is easy to determine if a given point on the surface is inside or outside of a given circumcircle we simply calculate distances in 3-sDace. The oDeration Point-inCircumcircle is a fundamental oper~tion for building Delaunay triangulations. A circumventer is always a point that is on the curved surface. This property is important since the vertices of a curved-surface mesh should lie up,on the surface and, in our meshing algorithm, new vertices are always circumventers. Finding a circumventer is basically operation Intersect-.Line-with-Surface. equivalent to the This operation

It would be nice if we could show an additional desirable property: that the number of triangles produced

277

triangle abc that has its center on our curved surface. Call the s~heres center p. Suppose d is outside of S. Now move p, the center of S, along the surface toward d while keeping b and con the boundary of S. Consider a plane through bc and perpendicular to the motion of the center. This plane cuts S into two pieces: on one side the sphere grows as the the center moves; on the other side the sphere shrinks. We need to ensure that (1) one side only shrinks and the other side only grows and (2) a is always on the side of the cutting plane where the sphere is shrinking. Both of these requirements can be met if we constrain the surface so that it does not bend too much in the region of our moving sphere. Thus, we have the following lemma. Lemma 5 Given a surface and given vertices a, b, c, and d on the surface where triangles abc, bed, abd, and acd all have well-defined circumventers, if the surface normals on the portion of the surface that is within the union o,f the circumcircles vary by less than ~ then the triangles have consistent circumcircles. Assuming we can start with an initial, relatively crude triangulation with the property that the surface does not bend too much near any pair of adjacent triangles, we can build something that can reasonably be called a Delaunay triangulation. Once we have a Delaunay triangulation, we also have a CDT since, intuitively, a CDT is a Delaunay triangulation in which certain edges cannot be flipped.

Figure circles

3: In a standard are consistent.

2D Delaunay

triangulation,

circum-

is available for all reasonable surface representation systems. (Unfortunately, this operation, although always available, is not necessarily an easy operation. For instance, when a surface is represented by parametric patches, intersecting the surface with a line is likely to require the use of an iterative method.) This definition of circumcircle stiU has some ditliculties. In particular, the line on which the sphere-centers lie may intersect the surface more than once or it may not intersect the surface at all. We claim that for triangulations that are reasonable reasonable in the sense that they are somewhat close to approximating their curved surface these difficulties either do not occur or can be easily resolved. Thus, this definition of circumcircle is useful if we have some way of creating a relatively crude, initial triangulation. This can be done by, for instance, intersecting the surface with a small-enough grid, or, in the case of surfaces represented by parametric patches, creating an initial triangulation by triangulating in the parameter space. At this point, we still need to show that the definition of circumcircle that we have suggested is an appropriate definition for building Delaunay triangulations. In particular, we need to prove a property that usually remains unstated for the regular 2D Delaunay triangulation: we need to show That that adjacent triangles have consistentcircumcircles. abc and bed, we need is, given edge bc with adj scent triangles to show that d is within the circumcircle of abc iff a is within the circumcircle of bed. In other words, we need to show that when we check the empty-circle property for a prospective Delaunay edge, we get the same answer regardless of which side we start from. As an example, consider the 2D version of this problem (see Figure 3). Let C be the circumcircle of triangle abc and suppose d is outside of C. Now move C toward d while (this will require that C keeping b and c on its boundary change size). By the time C touches d it is clear that a is no longer within C; thus, the circumcircle of bcd does not contain a and the two triangles have consistent circumcircles. For our 3D version, we let S represent the sphere through

THE CURVED-SURFACE ALGORITHM

MESHING

Just as in Algorithm A, triangles are graded according to two criteria: shape and size. Triangles are well-shaped if angles are greater than or equal to 30 degrees. It is important to note that our goal is to make meshes with triangles that are well-shaped as triangles in 3-space; thus, angles are measured in the plane of the triangles vertices. Triangles are well-sized if they paas a user-defined grading function, where the function is such that a small-enough, well-shaped triangle is always well-sized (see definition 1). For curvedsurface meshing, it is expected that the users size-grading function will take into account some measure of how well a given triangle matches the curved surface. ThE type of criterion is easily seen to satisfy Definition 1 as long as surface curvature is bounded away from infinity.

ALGORITHM
face.

B:

Meshing

for a region

on a curved

sur-

crude triangulation with the property 1. Build an initial that the surface normals in a region around each adjacent pair of triangles vary by less than ~. Use edge flips to make the initial triangulation into a curved-surface CDT. Some subdivision of triangles may be necessary.

2. Grade
triangle sized.

any triangles passes only

that are currently ungraded. A if it is both well-shaped and wellthe its

3. If all triangles

pass then Halt. Otherwise choose largest triangle that fails (call it A) and determine surface-circumventer (call it c).

4. Travel

across the CDT from any vertex of A toward c (any path wiIl work as long as we stay within the circumcircle of A) until we either run into a sourceedge or find the triangle containing c.

278

Figure

4: A curved

surface

mesh of a portion

of a wing.

5. If we found
the CDT.

the triangle containing Go to step 2.

c then

Insert

c into

6. If we ran into a source-edge then Split the edge and update the CDT. As in the 2D version of this algorithm, edges with lengths within a particular range can be split in thirds to simplify some proofs. In practice, splitting in half appears to always work. Note that, since boundaries may be curved, splitting in lhalf here means finding a point on the boundary (and on the surface) that is equidistant from each endpoint. Let 1 be the (straight-line) length of the new edges and consider the one or two new vertices produced by the Split operation. Delete each circumventer-vertex of the CDT that is closer than 1 (line-of-sight distance: an intervening source edge implies infinite distance) to a new vertex. Go to step 2. Asinthe 2Dcaae, ifthe algorithm halts then the resulting mesh consists entirely of triangles with desirable prop erties. For our curved surface version these properties can include that each triangle is a good approximation to the surface. See Figure 4 for an example surface mesh. To show the algorithm halts, we need to show that veryshort edges are not produced. In a sense, this is easier for curved surface meshes than for flat meshes, since the circumventer of a triangle is closest to the triangles vertices for flat circumventers (i.e., when the center of the sphere is in the same plane as the vertices); a curved-surface circumventer is usually farther away from the vertices. In the same way, splitting a curved boundary produces a midpoint that is farther from the endpoints than the midpoint of a straig,htIine boundary. Boundary splitting can be a bit complicated, since the new boundary may change the status of some nearby vertices, moving them from inside the region to outside the region or vice versa. This is resolved by extending the surfa,cenormal requirement to boundary-normals (vectors that are normal to the boundary and tangent to the surface). If the normals along a boundary do not vary by more than ~ t)hen any nearby vertices that change status are close enough that they are eliminated later in the splitting step.

opposite vertex of one triangle is within the circumcircle of the other. Even though we have created reasonable definitions of circumcircle and circumventer, it is still nontrivial to use these definitions in practice, since finding them may require iterative techniques (e.g., Newtons Method). Fortunately, experiment has shown our meshing technique to be extremely robust in the sense that, in practice, it appears we can Ask for better-shaped triangles. Angles between 35 and 100 degrees have caused no problems except in pathological cases specifically constructed to cause d~fficulties. Use approximate Delaunay triangulations. Instead of empty circumcircles, we can make-dc~ with nearlyempty circumcircles - nearly-empty in the sense that any vertices inside a nearly-empty circle are close to the circles boundary. Use approximate circumventers. If we use centers that are within distance fR of the true centers, where R is the circles radius and ~ is a fixedl fraction less than 1, then we can modify the meshing /algorithm and prove that all angles in the resulting mesh are greater than arcsin ~. For ~ = 0.5 this is about 14.5 degrees. In practice, it appears that the angle bound of 30 degrees can be maintained even whem~ approximate circumventers are used. The use of approximate circumventers is particularly effective at speeding up the generation of meshes for curved surfaces. We have developed 2D and curved-surface meshing techniques and shown that these techniques create meshes with the following properties: Boundaries and vertices specified by the user are respected. Triangles do not cross boundaries. Userspecified vertices are never eliminated.

30 degrees. All triangles have angles greater than (Smaller angles are required if we allow boundary angles less than 30 degrees.) Note that a triangles angles are always measured in the plane of the triangle.
Ihe user can control triangle size. Any kind of size criterion can be chosen as iong ss it ca~~ be achieved by making triangles smaller. Typical size criteria are based on error estimates (either intuitive or calculated) and on surface curvature. The curved-surface meshing aJgorithm is based on our development of a curved-surface Delaunay triangulation. Such a Delaunay triangulation is well-defined only if the given

SOME PRACTICAL AND CONCLUSIONS

IMPROVEMENTS

Our meshing technique is heavily baaed on our ability to maintain a Delaunay triangulation, flipping edges as necessary to recover Delaunay properties after the insertion c}f a vertex. This entails checking adjacent triangles to see if the

279

surface is smooth (surface normals vary by less than .;) in a region about each triangle. Fortunately, this restmction is not a severe handicap simple techniques can be used to generate a crude triangulation that is sufficiently dense for Delaunay operations (and our meshing technique) to be valid. For examples that we have tried, the initial crude triangulation typically consists of two to a few dozen triangles.

REFERENCES
[Rup92] J. Ruppert, A New and Simple Algorithm for Quality .2-Dimensional Mesh Generation, Report UCB/CSD 92/694, University of California, Berkeley, 1992. M. Bern and D. Eppstein, Mesh Generation and Optimal Triangulation, Computing in Euclidean Geometry, edited by F. K. Hwang and D.-Z. Du, World Scientific, 1992, to appear. Also appears as Tech Report CSL-92-1, Xerox PARC, March 1992. M. Bern, D. Eppstein, and J. R. Gilbert, Provably Good Mesh Generation, Proceedings of the 31st IEEE Symposium on the Foundations of Computer Science, 231-241, 1990. To appear in JCSS. L. P. Chew, Guaranteed-Quality Triangular Meshes, Department of Computer Science Tech 1989. Report TR 89-983, Cornell University,
S. A. Mitchell and S. A. Vavasis, Quality Mesh Generation in Three Dimensions, Proceedings of the Eighth Annual Symposium on Computational 212-221, ACM Press, 1992. Full verGeometry, sion in Department of Computer Science Tech Report TR 92-1267, Cornell University, 1992.

[BE92]

[BEG90]

[Che89]

[MV92]

[MS92]

E. A. Melissaratos and D. L. Souvaine, Coping with Inconsistencies: A New Approach to Produce Quality Triangulations of Polygons with Holes, Proceedings of the Eighth Annuai Symposium on Computational Geometry, 202-211, ACM Press, 1992.

280